test(deps): AUT-1177 validate MCore dev compatibility

.dev.commit

@@ -1 +1 @@
-0e6ac576fba8995fca541ceb6af5f35eb0e4f16e
+af91b6b5cc5aec01458b8e6d24ccb0f4717e49e3

.main.commit

@@ -1 +1 @@
-cd4afffa648426a959dc7cb1e24b5ce7d0c3ff54
+6513e3e23d6b5eda6a1c934990b15e804237732b

docker/common/import_check_skip.txt

@@ -35,3 +35,6 @@ rec2idx
 # Pytest plugin shipped by `hypothesis`; eagerly imports `pytest`, which is not
 # installed in the fw_base/fw_final runtime venvs (test-time-only dependency):
 _hypothesis_pytestplugin
+
+# MCore's setuptools build helper executes setup() when imported; it is not a runtime module:
+setup

examples/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+"""Megatron Bridge examples package."""

examples/conversion/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+"""Checkpoint conversion examples."""

examples/conversion/adapter/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+"""Adapter conversion examples."""

scripts/conversion/setup_conversion.py

@@ -149,12 +149,13 @@ def _build_executor(
     task_count = args.gpus_per_node if args.device == "gpu" else 1
     launcher = run.Torchrun() if args.executor == "local" and args.device == "gpu" else None
     if args.executor == "local":
-        return run.LocalExecutor(
-            nodes=1,
+        executor = run.LocalExecutor(
             ntasks_per_node=task_count,
             launcher=launcher,
             packager=run.Packager(),
         )
+        executor.nodes = 1
+        return executor
 
     gpu_kwargs = {}
     if args.device == "gpu" and not args.no_gpu_resource_request:

src/megatron/bridge/models/ernie_vl/modeling_ernie45_vl/model.py

@@ -167,6 +167,8 @@ def forward(
         position_ids: torch.Tensor,
         mrope_section,
         cp_group=None,
+        return_raw_freqs: bool = False,
+        packed_seq: bool = False,
     ) -> Tensor:
         """Compute ERNIE-style interleaved M-RoPE embeddings.
 
@@ -175,10 +177,14 @@ def forward(
             mrope_section: Ignored (kept for API compatibility). ERNIE uses
                 freq_allocation instead.
             cp_group: Context parallel group.
+            return_raw_freqs: Ignored because ERNIE uses interleaved rotary embeddings.
+            packed_seq: Whether the sequence uses THD packing. Packed sequences retain
+                their full rotary embedding for packed context-parallel handling.
 
         Returns:
             Tensor: RoPE embedding of shape [seq_len, batch, 1, head_dim].
         """
+        del return_raw_freqs
         device = self.inv_freq.device
         seq = position_ids.to(device=device, dtype=self.inv_freq.dtype)
         # seq: [3, bs, seq_len]
@@ -237,7 +243,7 @@ def forward(
 
         if cp_group is None:
             cp_group = self.cp_group
-        if cp_group is not None and cp_group.size() > 1:
+        if cp_group is not None and cp_group.size() > 1 and not packed_seq:
             emb = get_pos_emb_on_this_cp_rank(emb, 0, cp_group)
         return emb
 

src/megatron/bridge/models/gemma_vl/gemma3_vl_provider.py

@@ -61,8 +61,8 @@ class Gemma3VLModelProvider(Gemma3ModelProvider):
     def provide(self, pre_process=None, post_process=None, vp_stage=None) -> Gemma3VLModel:
         if not self.is_vision_language:
             raise ValueError("Gemma 3 VL requires is_vision_language=True.")
-        if self.attention_backend is not AttnBackend.fused:
-            raise ValueError("Gemma 3 VL requires the fused attention backend for attention bias.")
+        if self.attention_backend not in (AttnBackend.auto, AttnBackend.fused):
+            raise ValueError("Gemma 3 VL requires automatic or fused attention for attention bias.")
         if self.context_parallel_size != 1:
             raise ValueError("Gemma 3 VL does not support context parallelism with attention bias.")
 

src/megatron/bridge/models/megatron_mimo/megatron_mimo_provider.py

@@ -44,6 +44,28 @@
     from megatron.core.hyper_comm_grid import HyperCommGrid
 
 
+_GTP_REMAT_PROCESS_GROUP_ALIASES = (
+    "dp_cp_gtp_remat",
+    "expt_dp_gtp_remat",
+    "tp_ep_pp_with_egtp_remat",
+)
+
+
+def _get_gtp_remat_process_group_kwargs(
+    dp_cp_group: torch.distributed.ProcessGroup,
+    expt_dp_group: torch.distributed.ProcessGroup,
+    tp_ep_pp_group: torch.distributed.ProcessGroup,
+) -> Dict[str, torch.distributed.ProcessGroup]:
+    """Return GTP-remat aliases supported by the installed MCore contract."""
+    process_groups = {
+        "dp_cp_gtp_remat": dp_cp_group,
+        "expt_dp_gtp_remat": expt_dp_group,
+        "tp_ep_pp_with_egtp_remat": tp_ep_pp_group,
+    }
+    declared_fields = ProcessGroupCollection.__dataclass_fields__
+    return {name: process_groups[name] for name in _GTP_REMAT_PROCESS_GROUP_ALIASES if name in declared_fields}
+
+
 @dataclass
 class MegatronMIMOInfra:
     """MegatronMIMO infrastructure metadata (separate from model).
@@ -297,8 +319,13 @@ def _get_pg_collections_from_grids(
                 expt_dp_group = grid.get_pg(["expt_dp"], view=EXPERT_VIEW_NAME)
                 tp_ep_pp_group = grid.get_pg(["expt_tp", "ep", "pp"], view=EXPERT_VIEW_NAME)
 
-                # HyperCommGrid does not define GTP-remat axes, so the GTP-inclusive
+                # HyperCommGrid does not define GTP-remat axes, so supported GTP-inclusive
                 # process groups collapse to their existing non-GTP equivalents.
+                gtp_remat_process_groups = _get_gtp_remat_process_group_kwargs(
+                    dp_cp_group,
+                    expt_dp_group,
+                    tp_ep_pp_group,
+                )
                 pg_collections[module_name] = ProcessGroupCollection(
                     tp=grid.get_pg(["tp"]),
                     dp=grid.get_pg(["dp"]),
@@ -308,19 +335,17 @@ def _get_pg_collections_from_grids(
                     expt_tp=grid.get_pg(["expt_tp"], view=EXPERT_VIEW_NAME),
                     expt_dp=expt_dp_group,
                     dp_cp=dp_cp_group,
-                    dp_cp_gtp_remat=dp_cp_group,
-                    expt_dp_gtp_remat=expt_dp_group,
                     intra_dp_cp=dp_cp_group,
                     tp_cp=grid.get_pg(["tp", "cp"]),
                     tp_dp_cp=grid.get_pg(["tp", "dp", "cp"]),
                     mp=grid.get_pg(["tp", "pp"]),
                     tp_ep=grid.get_pg(["expt_tp", "ep"], view=EXPERT_VIEW_NAME),
                     tp_ep_pp=tp_ep_pp_group,
-                    tp_ep_pp_with_egtp_remat=tp_ep_pp_group,
                     intra_expt_dp=expt_dp_group,
                     intra_dist_opt=grid.get_pg(["tp", "cp", "dp", "pp"]),
                     pos_embd=pos_embd_pg if first_stage else None,
                     embd=embd_pg if (first_stage or last_stage) else None,
+                    **gtp_remat_process_groups,
                 )
             else:
                 pg_collections[module_name] = None

src/megatron/bridge/recipes/gemma3_vl/h100/gemma3_vl.py

@@ -70,7 +70,7 @@ def gemma3_vl_4b_sft_1gpu_h100_bf16_config() -> ConfigContainer:
     cfg.model.cuda_graph_warmup_steps = 3
 
     # Kernel selections
-    cfg.model.attention_backend = AttnBackend.fused
+    cfg.model.attention_backend = AttnBackend.auto
     cfg.model.cross_entropy_loss_fusion = True
     cfg.model.cross_entropy_fusion_impl = "native"
 
@@ -182,7 +182,7 @@ def gemma3_vl_12b_sft_4gpu_h100_bf16_config() -> ConfigContainer:
     cfg.model.cuda_graph_warmup_steps = 3
 
     # Kernel selections
-    cfg.model.attention_backend = AttnBackend.fused
+    cfg.model.attention_backend = AttnBackend.auto
     cfg.model.cross_entropy_loss_fusion = True
     cfg.model.cross_entropy_fusion_impl = "native"
 
@@ -294,7 +294,7 @@ def gemma3_vl_27b_sft_16gpu_h100_bf16_config() -> ConfigContainer:
     cfg.model.cuda_graph_warmup_steps = 3
 
     # Kernel selections
-    cfg.model.attention_backend = AttnBackend.fused
+    cfg.model.attention_backend = AttnBackend.auto
     cfg.model.cross_entropy_loss_fusion = True
     cfg.model.cross_entropy_fusion_impl = "native"
 
@@ -415,7 +415,7 @@ def gemma3_vl_4b_peft_1gpu_h100_bf16_config(peft_scheme: str | PEFT = "lora") ->
     cfg.model.cuda_graph_warmup_steps = 3
 
     # Kernel selections
-    cfg.model.attention_backend = AttnBackend.fused
+    cfg.model.attention_backend = AttnBackend.auto
     cfg.model.cross_entropy_loss_fusion = True
     cfg.model.cross_entropy_fusion_impl = "native"
 
@@ -536,7 +536,7 @@ def gemma3_vl_12b_peft_1gpu_h100_bf16_config(peft_scheme: str | PEFT = "lora") -
     cfg.model.cuda_graph_warmup_steps = 3
 
     # Kernel selections
-    cfg.model.attention_backend = AttnBackend.fused
+    cfg.model.attention_backend = AttnBackend.auto
     cfg.model.cross_entropy_loss_fusion = True
     cfg.model.cross_entropy_fusion_impl = "native"
 
@@ -657,7 +657,7 @@ def gemma3_vl_27b_peft_4gpu_h100_bf16_config(peft_scheme: str | PEFT = "lora") -
     cfg.model.cuda_graph_warmup_steps = 3
 
     # Kernel selections
-    cfg.model.attention_backend = AttnBackend.fused
+    cfg.model.attention_backend = AttnBackend.auto
     cfg.model.cross_entropy_loss_fusion = True
     cfg.model.cross_entropy_fusion_impl = "native"
 

tests/unit_tests/conversion/launcher/test_setup_conversion.py

@@ -257,7 +257,13 @@ def test_local_cpu_executor_uses_one_process_without_launcher():
     module = _load_setup_conversion_module()
     captured = {}
     module.run.Packager = lambda: "packager"
-    module.run.LocalExecutor = lambda **kwargs: captured.update(kwargs) or types.SimpleNamespace(**kwargs)
+
+    def local_executor(**kwargs):
+        assert "nodes" not in kwargs
+        captured.update(kwargs)
+        return types.SimpleNamespace(**kwargs)
+
+    module.run.LocalExecutor = local_executor
     args = _parse(module)
 
     executor = module._build_executor(args, [], [])

tests/unit_tests/models/ernie_vl/test_ernie45_vl_bridge.py

@@ -33,6 +33,7 @@
     ErnieMultiTypeMoE,
     MultiTypeMoeSubmodules,
 )
+from megatron.bridge.models.ernie_vl.modeling_ernie45_vl.model import ErnieMultimodalRotaryEmbedding
 
 
 def _make_vision_config():
@@ -469,6 +470,52 @@ def test_offset_row_parallel_mapping_resolve(self):
         assert "64" in resolved.hf_param
 
 
+class TestErnieMultimodalRotaryEmbedding:
+    """Test ERNIE multimodal rotary context-parallel handling."""
+
+    @pytest.fixture
+    def rotary_embedding(self):
+        """Create a CPU rotary embedding without invoking the CUDA constructor."""
+        rotary_embedding = ErnieMultimodalRotaryEmbedding.__new__(ErnieMultimodalRotaryEmbedding)
+        torch.nn.Module.__init__(rotary_embedding)
+        rotary_embedding.inv_freq = torch.tensor([1.0, 0.5])
+        rotary_embedding.seq_len_interpolation_factor = None
+        rotary_embedding.freq_allocation = 0
+        rotary_embedding.cp_group = None
+        return rotary_embedding
+
+    @pytest.fixture
+    def position_ids(self):
+        """Create distinct temporal, height, and width positions."""
+        return torch.arange(12).reshape(3, 1, 4)
+
+    def test_unpacked_sequence_slices_context_parallel_embedding(self, rotary_embedding, position_ids):
+        cp_group = Mock()
+        cp_group.size.return_value = 2
+
+        with patch(
+            "megatron.bridge.models.ernie_vl.modeling_ernie45_vl.model.get_pos_emb_on_this_cp_rank",
+            side_effect=lambda embedding, *_: embedding[:2],
+        ) as get_cp_embedding:
+            output = rotary_embedding(position_ids, [1, 1, 0], cp_group=cp_group)
+
+        assert output.shape == (2, 1, 1, 4)
+        get_cp_embedding.assert_called_once()
+        assert get_cp_embedding.call_args.args[1:] == (0, cp_group)
+
+    def test_packed_sequence_retains_full_context_parallel_embedding(self, rotary_embedding, position_ids):
+        cp_group = Mock()
+        cp_group.size.return_value = 2
+
+        with patch(
+            "megatron.bridge.models.ernie_vl.modeling_ernie45_vl.model.get_pos_emb_on_this_cp_rank"
+        ) as get_cp_embedding:
+            output = rotary_embedding(position_ids, [1, 1, 0], cp_group=cp_group, packed_seq=True)
+
+        assert output.shape == (4, 1, 1, 4)
+        get_cp_embedding.assert_not_called()
+
+
 class TestErnie45VLModelProvider:
     """Test Ernie45VLModelProvider class."""
 

tests/unit_tests/models/gemma/test_gemma4_modeling.py

@@ -923,7 +923,7 @@ def test_moe_global_rotary_matches_hf_proportional_coordinate_layout(self):
         hidden_states = torch.arange(1, head_dim + 1, dtype=torch.float32).view(1, 1, 1, head_dim)
         freqs = rotary.get_freqs_non_repeated(1, offset=position)
         freqs = torch.cat((freqs, freqs), dim=-1)[:, None, None, :]
-        config = SimpleNamespace(apply_rope_fusion=False, rotary_interleaved=False)
+        config = SimpleNamespace(apply_rope_fusion=False, rotary_interleaved=False, mrope_section=None)
         actual = apply_rotary_pos_emb(hidden_states, freqs, config, cp_group=object())
 
         expected = hidden_states.clone()

tests/unit_tests/models/gemma_vl/test_gemma3_vl_provider.py

@@ -213,6 +213,20 @@ def test_gemma3_vl_rejects_causal_override(self):
         with pytest.raises(ValueError, match="requires is_vision_language=True"):
             provider.provide()
 
+    def test_gemma3_vl_accepts_auto_attention_override(self):
+        """Test that automatic backend selection preserves the VL attention bias."""
+        provider = Gemma3VLModelProvider(
+            num_layers=28,
+            hidden_size=2560,
+            num_attention_heads=10,
+        )
+        provider.attention_backend = AttnBackend.auto
+
+        with patch("megatron.bridge.models.gemma_vl.gemma3_vl_provider.Gemma3VLModel") as model:
+            provider.provide()
+
+        model.assert_called_once()
+
     def test_gemma3_vl_rejects_flash_attention_override(self):
         """Test that runtime overrides cannot select a backend that drops the VL mask."""
         provider = Gemma3VLModelProvider(
@@ -222,7 +236,7 @@ def test_gemma3_vl_rejects_flash_attention_override(self):
         )
         provider.attention_backend = AttnBackend.flash
 
-        with pytest.raises(ValueError, match="requires the fused attention backend"):
+        with pytest.raises(ValueError, match="requires automatic or fused attention"):
             provider.provide()
 
     def test_gemma3_vl_custom_vision_projector_config(self):

tests/unit_tests/models/megatron_mimo/test_megatron_mimo_provider.py

@@ -13,6 +13,7 @@
     MegatronMIMOParallelismConfig,
     ModuleParallelismConfig,
 )
+from megatron.bridge.models.megatron_mimo.megatron_mimo_provider import _get_gtp_remat_process_group_kwargs
 
 
 class FakeStandardProvider:
@@ -683,6 +684,34 @@ def test_is_pp_last_stage_none_group(self):
 class TestProcessGroupCollectionWithEmbeddingGroups:
     """Test that ProcessGroupCollection includes embedding groups."""
 
+    @patch(
+        "megatron.bridge.models.megatron_mimo.megatron_mimo_provider.ProcessGroupCollection.__dataclass_fields__",
+        {"tp": Mock()},
+    )
+    def test_gtp_remat_aliases_omitted_for_older_mcore_contract(self):
+        """Test that aliases absent from the MCore contract are not passed to its constructor."""
+        assert _get_gtp_remat_process_group_kwargs(Mock(), Mock(), Mock()) == {}
+
+    @patch(
+        "megatron.bridge.models.megatron_mimo.megatron_mimo_provider.ProcessGroupCollection.__dataclass_fields__",
+        {
+            "dp_cp_gtp_remat": Mock(),
+            "expt_dp_gtp_remat": Mock(),
+            "tp_ep_pp_with_egtp_remat": Mock(),
+        },
+    )
+    def test_gtp_remat_aliases_preserved_for_current_mcore_contract(self):
+        """Test that current MCore receives all collapsed GTP-remat aliases."""
+        dp_cp_group = Mock()
+        expt_dp_group = Mock()
+        tp_ep_pp_group = Mock()
+
+        assert _get_gtp_remat_process_group_kwargs(dp_cp_group, expt_dp_group, tp_ep_pp_group) == {
+            "dp_cp_gtp_remat": dp_cp_group,
+            "expt_dp_gtp_remat": expt_dp_group,
+            "tp_ep_pp_with_egtp_remat": tp_ep_pp_group,
+        }
+
     @patch("megatron.bridge.models.megatron_mimo.megatron_mimo_provider.is_pp_last_stage")
     @patch("megatron.bridge.models.megatron_mimo.megatron_mimo_provider.is_pp_first_stage")
     @patch("megatron.bridge.models.megatron_mimo.megatron_mimo_provider.populate_embedding_and_position_groups")
@@ -835,15 +864,17 @@ def test_pg_collection_preserves_dense_and_expert_view_contracts(
         assert pgc.ep == mock_ep
         assert pgc.expt_tp == mock_expt_tp
         assert pgc.expt_dp == mock_expt_dp
-        assert pgc.expt_dp_gtp_remat == mock_expt_dp
         assert pgc.dp_cp == mock_dp_cp
-        assert pgc.dp_cp_gtp_remat == mock_dp_cp
         assert pgc.intra_dp_cp == mock_dp_cp
+        if "expt_dp_gtp_remat" in pgc.__dataclass_fields__:
+            assert pgc.expt_dp_gtp_remat == mock_expt_dp
+            assert pgc.dp_cp_gtp_remat == mock_dp_cp
         assert pgc.tp_cp == mock_tp_cp
         assert pgc.tp_dp_cp == mock_tp_dp_cp
         assert pgc.mp == mock_mp
         assert pgc.tp_ep == mock_tp_ep
         assert pgc.tp_ep_pp == mock_tp_ep_pp
-        assert pgc.tp_ep_pp_with_egtp_remat == mock_tp_ep_pp
+        if "tp_ep_pp_with_egtp_remat" in pgc.__dataclass_fields__:
+            assert pgc.tp_ep_pp_with_egtp_remat == mock_tp_ep_pp
         assert pgc.intra_expt_dp == mock_expt_dp
         assert pgc.intra_dist_opt == mock_intra_dist_opt

tests/unit_tests/recipes/test_gemma3_vl_recipes.py

@@ -97,7 +97,7 @@ def _assert_basic_config(cfg):
     assert cfg.tokenizer is not None
     assert cfg.checkpoint is not None
     assert cfg.rng is not None
-    assert cfg.model.attention_backend is AttnBackend.fused
+    assert cfg.model.attention_backend is AttnBackend.auto
     assert cfg.dataset.enable_in_batch_packing is False
 
     assert cfg.train.global_batch_size >= 1